Академический Документы
Профессиональный Документы
Культура Документы
doi: 10.1093/rof/rfp018
Advance Access publication: 4 October 2009
Abstract. We test the limits of arbitrage argument for the survival of irrationality-induced financial
anomalies by sorting securities on their individual residual variability as a proxy for idiosyncratic
risk – a commonly asserted limit to arbitrage – and comparing the strength of anomalous returns
in low versus high residual variability portfolios. We find no support for the limits of arbitrage
argument to explain undervaluation anomalies (small value stocks, value stocks generally, recent
winners, and positive earnings surprises) but strong support for the limits of arbitrage argument to
explain overvaluation anomalies (small growth stocks, growth stocks generally, recent losers, and
negative earnings surprises). Other tests also fail to support the limits of arbitrage argument for the
survival of overvaluation anomalies and suggest that at least some of the factor premiums for size,
book-to-market, and momentum are unrelated to irrationality protected by limits to arbitrage.
1. Introduction
Empirical asset pricing tests of the predictions of the Sharpe-Lintner CAPM often
result in model falsification. Small stocks earn returns that are higher than predicted
(see Banz, 1981), as do recent winners (see, e.g., Chan et al., 1996), value stocks
(see, e.g., Lakonishok et al., 1994), and stocks of companies with positive earnings
surprises (see, e.g., Ball and Brown, 1968; Bernard and Thomas, 1990). Growth
We thank Ray Ball, Nick Barberis, Zahi Ben-David, Peter Bossaerts (the editor), Michael Brandt,
Markus Brunnermeier, George Constantinides, Campbell Harvey, Dong Hong, Ron Kaniel, Reuven
Lehavy, Jon Lewellen, Mark Loewenstein (a discussant), Toby Moskowitz, Doron Nissim, Per Olsson,
Matthew Rothman, Darren Roulstone, Ronnie Sadka, Richard Thaler, Mohan Venkatachalam, Mitch
Warachka, two anonymous referees, and seminar participants at the Seventh Maryland Finance
Symposium, the Securities and Exchange Commission Office of Economic Analysis, Duke University,
Hebrew University, Interdisciplinary Center, Herzlyia, Israel, Northwestern University and Vanderbilt
University for helpful comments. Please address correspondence to Brav at Fuqua School of Business,
Duke University, Box 90120, Durham, North Carolina 27708–0120, email: brav@duke.edu. Heaton
acknowledges that the opinions expressed here are his own, and do not reflect the position of Bartlit
Beck Herman Palenchar & Scott LLP or its attorneys. Li acknowledges financial support from the
Social Sciences and Humanities Research Council of Canada.
C The Authors 2009. Published by Oxford University Press [on behalf of the European Finance Association].
All rights reserved. For Permissions, please email: journals.permissions@oxfordjournals.org
158 A. BRAV ET AL.
stocks, recent losers, and negative earnings surprises earn returns that are lower
than predicted (see, e.g., Ritter, 1991; Barberis and Huang, 2008; Ball and Brown,
1968; Bernard and Thomas, 1990; Chan et al., 1996).
Researchers in behavioral finance argue that asset pricing anomalies result
from the influence of unmodeled irrational behavior on security prices (see, for
example, the extended discussion in Barberis and Thaler (2003) of the above anoma-
lies). The behavioral claim is controversial. First, rational behavior is just one of
several assumptions used to derive the tested asset pricing models. Model falsifi-
cation might result from the failure of an assumption other than the assumption of
rationality (see Cross, 1982; Lowenstein and Willard, 2006; Fama, 1991). Second,
the behavioral claim implies that rational arbitrageurs cannot exploit the irrational
investors and drive irrationality-induced price deviations to zero (see Friedman,
1953; Fama, 1965) or near-zero (see Grossman and Stiglitz, 1980) levels. We call
this the “arbitrage objection.”
To rebut the arbitrage objection, some researchers have argued that arbitrage is
more difficult than recognized by those offering the arbitrage objection. Barberis
and Thaler (2003) survey the literature and identify three sources of limits to
arbitrage: idiosyncratic risk, noise trader momentum risk, and implementation costs.
Idiosyncratic risk is proposed as a limit to arbitrage in several papers.1 In their
paper, “The Limits of Arbitrage,” Shleifer and Vishny (1997) explain their model
as follows:
To specialized arbitrageurs, both systematic and idiosyncratic volatility matters. In fact, id-
iosyncratic volatility probably matters more, since it cannot be hedged and arbitrageurs are
not diversified. . . . In our model . . . stocks are not rationally priced, and idiosyncratic risk
deters arbitrage. . . . For a given noise trading process, volatile securities will exhibit greater
mispricing and a higher average return to arbitrage.
Noise trader momentum risk, according to Barberis and Thaler (2003), is the risk
that irrational beliefs get worse in the direction already distorting prices:
Noise trader risk . . . is the risk that the mispricing being exploited by the arbitrageur worsens in
the short run. Even if General Motors is a perfect substitute for Ford, the arbitrageur still faces
1
See also Pontiff (1996); Ali et al. (2003); Mendenhall (2004); Wurgler and Zhuravskaya (2002);
Mashruwala et al. (2006); and Pontiff (2006).
THE LIMITS OF THE LIMITS OF ARBITRAGE 159
the risk that the pessimistic investors causing Ford to be undervalued in the first place become
even more pessimistic, lower its price even further. Once one has granted the possibility that
a security’s price can be different from its fundamental value, then one must also grant the
possibility that future price movements will increase the divergence.
Finally, Barberis and Thaler (2003) posit that “[w]ell understood transaction costs
such as commissions, bid-ask spreads and price impact can make it less attractive
to exploit a mispricing.”
In this paper, we report the results of a set of tests designed to test the limits of ar-
bitrage argument. Under the hypothesis that greater idiosyncratic volatility protects
the existence of greater levels of mispricing, our first set of tests sorts securities on
the size of their residual variability from the Fama and French four-factor model
(as a measure of idiosyncratic risk) and ask whether financial anomalies increase
in magnitude with the amount of residual variability. Under the hypothesis that
momentum in the same direction protects greater levels of mispricing, in our sec-
ond set of tests we form portfolios of value stocks – which, according to some
researchers, earn abnormal positive returns because irrational investors extrapolate
past bad performance into the future (see, e.g., Lakonishok et al., 1994) – and
sort them into those exhibiting negative momentum and those exhibiting positive
momentum, testing the prediction that arbitrage against loser-value stocks is riskier
and thus should deter arbitrage and allow for the survival of higher abnormal returns
to value stocks.
Our first set of tests reject the limits of arbitrage argument for positive return
anomalies – small value stocks, value stocks generally, recent winners, and posi-
tive earnings surprises – since in each case we do not find the relation predicted
by the limits of arbitrage argument. Instead, these anomalies are stronger when
idiosyncratic risk is low. These results are robust to alternative specifications of
idiosyncratic risks and transactions costs. In fact, the high positive correlations be-
tween total and residual volatility and other generally accepted measures of limits
of arbitrage like the extent of institutional holding, analyst coverage, and stock price
level suggest that it is unlikely that any alternative measures of idiosyncratic risk
or transactions costs could change our inferences. To reverse our inferences, alter-
native measures would have to be both different and move in the opposite direction
from current proxies. It is difficult to imagine such alternative measures retaining
plausible interpretation as limits of arbitrage. The evidence is not consistent with
the limits of arbitrage argument for the survival of irrationality-induced underval-
uation of small value stocks, value stocks generally, recent winners, and positive
earnings surprises. However, our first set of tests does strongly support the limits of
arbitrage argument for the abnormal returns to small growth stocks, growth stocks
generally, recent losers, and negative earnings surprises. Negative returns to these
anomalies occur only among stocks with high residual variability from the Fama
and French four-factor model.
160 A. BRAV ET AL.
Our second set of tests rejects the limits of arbitrage argument for both under-
valuation and overvaluation anomalies. In those tests we ask whether noise trader
momentum risk might allow more mispricing of value and growth firms when re-
cent returns have been bad (for value firms) or good (for growth firms). The idea
is simple: if value firms are mispriced due to excessive extrapolation of bad perfor-
mance, those firms that have had very recent poor performance have momentum
in the direction of the mispricing and should be riskier to bet against. The same
holds true for overpriced growth firms that have been doing well recently. But we
find the opposite to be true: value firms that are recent losers have lower returns
than value firms that are recent winners, and growth firms that are recent losers
do worse than growth firms that are recent winners. Both results are inconsistent
with the behavioral null provided by the noise trader momentum risk version of the
limits of arbitrage argument.
Finally, we ask whether there is a factor premium for size, book-to-market, and
momentum in portfolios comprised of the lowest residual variability stocks. The
lowest residual variability stocks not only have low idiosyncratic risk but also
have high median prices, high institutional holdings, a larger number of analysts,
considerable liquidity, high divided yields, and comprise a large part of the market
capitalization of the entire market. It is difficult to argue that limits of arbitrage are
meaningful in these stocks. If the factor premiums for size, book-to-market, and
momentum were driven entirely by irrationality then we would expect very little
covariation with these factors in the lowest residual variability portfolios. Instead,
we find a strong role for these factors in explaining the returns to zero cost portfolios
of the lowest residual variability stocks.
The rest of the paper is organized as follows. Section 2 discusses earlier attempts
to test the limits of arbitrage argument. We present the data and methodology in
Section 3. In Section 4, we report the first set of results linking limits to arbitrage and
average returns to size, value, momentum, and earnings surprise based strategies.
Section 5 provides an additional test based on the profitability of investing in
both value and momentum strategies. In Section 6 we ask whether size, value and
momentum premia exist even in low limits to arbitrage environment. We conclude
the paper in Section 7.
We reach opposite conclusions than other recent papers that purport to confirm
limits of arbitrage explanations of financial anomalies that we reject here, in-
cluding Ali et al. (2003), Zhang (2006), and Mendenhall (2004). We find that
prior studies, by equal-weighting rather than value-weighting returns, by failing to
THE LIMITS OF THE LIMITS OF ARBITRAGE 161
in absolute value as one holds portfolios of high residual variability stocks, a proxy
for limits to arbitrage. His evidence is inconsistent with our finding that the pos-
itive drift is actually confined to low limit to arbitrage environments. However,
while we design our tests around a simple passive and implementable calendar
time trading strategy with appropriate adjustment for common risk factors in re-
turns, Mendenhall’s design considers only non-implementable event-time analysis,
in which returns are equally weighted and risk or factor adjustment is made only
with respect to size matched portfolios.2 Of course, one might argue that an anomaly
identified by a non-implementable trading strategy necessarily invokes the limits of
arbitrage argument because a trading strategy that is non-implementable is necessar-
ily a limit to arbitrage. This calls into question whether the anomaly is economically
interesting since it does not exist in a tradeable form. Moreover, Mendenhall uses
a market model to form estimates of residual volatility, leaving common factors in
returns due to size, book to market and momentum, which are therefore included in
his measure of residual variability. Sorting on his market model residual volatility
results in a sort on these firm characteristics. As a result, his control for size related
differences in average returns is incomplete.
We obtain stock returns from the Center for Research in Security Prices for all firms
traded on the NYSE, AMEX, and NASDAQ, subject to the restriction that they
have CRSP common share codes equal to 10 or 11, from July 1958 to December
2007. Accounting data is obtained from Compustat. We measure the magnitude
of specific anomalies with average excess returns, alpha estimates, and Sharpe
Ratios (annualized excess returns per unit of standard deviation) for small stocks
(including small growth stocks and small value stocks), recent winner stocks, recent
2
The use of calendar-time regressions to measure abnormal performance has been questioned by
Loughran and Ritter (2000). They argue that the implementation of calendar-time regressions can
result in low power to reject a null model in tests of market efficiency. One concern is the choice of
equal versus value weighting of portfolio returns. If mispricing is likely to occur in hard to arbitrage
small firms then a weighting scheme that tilts away from these firms will result in low power. We share
this concern and thus report, for each undervaluation and overvaluation anomaly, tests that restrict the
sample to the subset of small firms. Loughran and Ritter (2000) also argue that benchmark factors
might be “contaminated” by the test assets that are to be priced. For example, the inclusion of the
book to market factor, HML, in the Fama and French three-factor model may lead to low power since
many of the issuing firms that they try to price are included in this factor. It is unclear, however, why
this concern applies in the context of our tests. As we show in Section II, the calendar-time setup
affords the power needed to support our conclusions. At any rate, our conclusions do not rest on
the three or four -factor models as our tests are benchmarked purposely against the CAPM precisely
because we want to measure abnormal returns against a null model that behavioral finance has no
objections to.
THE LIMITS OF THE LIMITS OF ARBITRAGE 163
loser stocks, value stocks, growth stocks, and stocks subject to both positive and
negative extreme earnings surprises. For each characteristic-based anomaly (e.g.,
small stocks) we estimate, beginning in July of 1963, for each firm, a four-factor
regression using monthly return data from the preceding five years. We use a
minimum of 36 monthly returns to estimate the regressions. Firms are allocated
to quartile portfolios based on the magnitude of the estimated residual standard
deviation and we value weight firm returns for the ensuing three months. We
require a minimum of 50 firms in each portfolio. This portfolio formation is
repeated quarterly through December 2007.3 We estimate idiosyncratic risk from
a four-factor asset-pricing model including the Fama and French RMRF, SMB,
and HML factors and a momentum factor, MOM. We obtain these factor returns
and monthly risk-free rates (which we use to calculate excess returns) from Ken
French’s web site at Dartmouth College.4
Panel A of Table I presents results for small stocks. Small stocks are those traded on
NYSE, AMEX, and NASDAQ with CRSP common share codes 10 or 11 that fall in
the bottom size quintile using NYSE market capitalization quintile breakpoints. We
allocate these stocks into quartile sorts based on the magnitude of their estimated
residual variability. Q1-Low (Q4-High) is a portfolio holding firms with lowest
(highest) estimated residual variability.
An immediately interesting result – one that appears throughout our tests – is the
apparent existence of large amounts of covariation among stocks with high residual
variability. The R-squared of the portfolio of the highest residual variability stocks
falls dramatically to 73.51 from 88.06 for the lowest residual variability stocks.
It is worthwhile noting that this covariation, if it is not hedgeable, makes betting
against high residual variability stocks even riskier since the idiosyncratic risk will
not even disappear in large portfolios of potentially mispriced securities.
High returns to small stocks occur only in the lowest idiosyncratic risk quartiles,
and not at all in the highest quartile, the opposite of the relation predicted by the
limits of arbitrage argument. When we perform CAPM regressions on these port-
folio returns, there are no anomalous returns relative to the CAPM for small stocks
in high idiosyncratic risk portfolios. The CAPM prices the portfolio of the highest
residual variability stocks with an insignificant negative alpha of 34 basis points
3
None of our conclusions change if portfolios are rebalanced every six or twelve months, although
the overall profitability of the momentum strategy (both winners and losers) declines.
4
http://mba.tuck.dartmouth.edu/pages/faculty/ken.french/data_library.html.
164 A. BRAV ET AL.
Table I. Continued
Table I. Continued
per month. The lowest idiosyncratic risk portfolio, however, earns a statistically
significant positive 54 basis points per month. The same result hold for returns
relative to the Fama and French four-factor model. The anomalous returns to the
smallest stocks are statistically significant and positive for the lowest idiosyncratic
risk portfolio (0.19; t-statistic of 2.59) while the highest idiosyncratic risk quartile
earns statistically significant negative returns (−0.59; t-statistic −2.70).
The annualized Sharpe Ratio increases monotonically from the highest idiosyn-
cratic risk quartile to the lowest, contrary to the predictions of the limits of arbitrage
argument. The Sharpe Ratio of the lowest idiosyncratic risk quartile is 0.72, while
the Sharpe Ratio of the highest idiosyncratic risk quartile is 0.15. The difference
between these two Sharpe Ratios is both a return and a standard deviation effect.
The highest idiosyncratic risk portfolio has the lowest portfolio average monthly
excess return of 0.38 and the highest portfolio monthly standard deviation of 9.10.
Portfolio standard deviations decrease monotonically from the highest idiosyncratic
risk quartile to the lowest. Since we form these portfolios using individual security
residual variance, this tells us that high residual variance securities tend also to have
exposures to common factors in returns that do not wash out in large portfolios.
That is, firms with high residual variance tend to share high common variation in
return. This is apparent from the pattern of four factor loadings. Market loadings
increase monotonically from 0.77 in the lowest idiosyncratic risk quartile to 1.30 in
the highest idiosyncratic risk quartile. SMB loadings increase monotonically from
0.69 to 1.31 from the lowest to the highest idiosyncratic risk quartile while the
loadings on HML decrease monotonically from 0.58 to 0.14. Loadings on the mo-
mentum factor exhibit a small increase from the lowest to the highest idiosyncratic
risk quartiles. This pattern is apparent in all of our results: high residual variability
securities have high market betas and high SMB loadings.
We next explore the evidence with respect to stock price level, institutional
ownership, analyst following, liquidity, and market capitalization. ‘Median stock
price level’ provides information on the price level of securities allocated to the
THE LIMITS OF THE LIMITS OF ARBITRAGE 167
portfolios. Each quarter, we record the median stock price of firms allocated to a
given portfolio as of the preceding month and report the average across the full
sample period. For small stocks, median prices fall from the lowest to the highest
idiosyncratic risk quartiles. The median price of the lowest idiosyncratic risk quartile
is $15.17, decreasing to $3.93 in the highest idiosyncratic risk quartile. We obtain
institutional holdings from Thomson Financial’s CDA/Spectrum Institutional (13f)
Holdings data. ‘Institutional Holding’ is the percent of shares held by institutions
calculated by averaging the percent shares held in the pre-formation month and
then averaged across time. For small stocks, institutional holding in the lowest
idiosyncratic risk quartile is nearly twice that in the high limits portfolio. Number
of analysts following a firm is obtained from IBES. We calculate the average
number of analysts per firm in the quarter post-formation and then average across
firms and across time. For small stocks, there is no clear pattern in the number of
analysts.
Illiquid stocks may be harder to arbitrage. We calculate two measures that proxy
for illiquidity. The first, ‘Percent dollar volume’, is calculated by first obtaining the
sum of dollar volume traded of all firms in a given portfolio. Then, for each quartile
portfolio, we calculate the fraction of dollar volume relative to the total dollar
volume summed over all four portfolios. This percentage is then averaged across
time. Our second measure, ‘Amihud Measure’, is the Amihud (2002) illiquidity
measure. We calculate a monthly measure by taking the ratio of absolute value
of daily return scaled by daily dollar volume, averaged within a month. For each
portfolio we report the average of this measure across the allocated securities in the
pre-formation quarter and then across time. For small stocks there is no evidence
of a monotonic sort on percent dollar volume. However, illiquidity, proxied by
the Amihud measure, increases monotonically from the lowest to the highest limit
portfolios.
‘Campbell IV’ is a monthly estimate of stock idiosyncratic volatility calculated
based on daily data for each firm as in Campbell et al. (2001). High residual
variability firms are more volatile under this measure as well. Dividend yield is
calculated as the difference between the buy and hold return including dividends in
the 12 months post-formation and the buy and hold return, over the same period,
excluding dividends. We then average these yields across firms in a given portfolio
and then across time. Dividend yield falls from low residual variability to high
residual variability. Pontiff (1996) explains that dividend yield may lower arbitrage
costs by shortening the duration of a position in a mispriced security. Hoberg and
Prabhala (2009) also find that idiosyncratic volatility is negatively correlated with
dividend yield.5 Finally, percent of market cap is an estimate of the wealth invested
5
An alternative proxy for limited arbitrage in the context of anomalies involving overvaluation is
the cost of shorting. While we do not have access to such data, the evidence in Geczy et al. (2002) is
168 A. BRAV ET AL.
Zellner and Siow (1979) who approximate the posterior odds for the test that the
mean is greater than zero versus less than zero as (prior odds)×(F(t)/F(-t), where F is
the cumulative standard normal distribution, and the prior probability distributions
are Cauchy centered on zero and diffuse on the standard deviation. With even prior
odds (that is, a prior belief that it is equally likely that the high residual portfolio is
lower than the low residual variability portfolio as it is that the opposite is true) the
posterior odds for the hypothesis that the high residual variability portfolio earns
more than the low residual variability portfolio are vanishingly small at 7.35 E-10.
In other words, the odds on the limits of arbitrage argument for small growth stocks
are about 1.4 billion to 1 in favor of that hypothesis.
Panel C of Table I presents results for small value stocks. These are firms which
are classified in the sorting month as falling in the bottom size quintile and the top
book to market quintile. Inconsistent with limits of arbitrage explanations of high
returns to small value stocks, the highest CAPM and four factor alphas occur in
the lowest residual variability portfolio subset of the small value stocks. Similarly,
the Sharpe Ratio of the lowest residual variability portfolio is much higher than the
Sharpe Ratio of the highest residual variability portfolio (0.84 versus 0.52). This
is largely a standard deviation effect. While the low residual variability portfolio
of small value stocks earns 4 basis points less per month relative to the high limits
portfolio, the standard deviation of the highest residual variability portfolio is 60%
higher (8.06 versus 4.80). The CAPM alpha is larger for low limits small value
stocks than high limits small value stocks. It is worth noting that not even the
170 A. BRAV ET AL.
four-factor model can explain low residual variability small value stocks. The four-
factor alpha is 0.32 with a t-statistic of 3.60. There is no four-factor anomaly in
high residual variability small value stocks.
Table II presents the results of zero cost portfolio CAPM regressions that are
long the high residual variability small value portfolio and short the low residual
variability small value portfolio. Under the null of the limits of arbitrage argument,
the sign on the small value regression alpha should be positive, since the positive
alpha generated by small value stocks should be strongest in the highest residual
variability portfolio if the anomaly is strongest where limits to arbitrage are highest.
We do not find support for the limits of arbitrage argument for small value stocks.
The CAPM alpha of the zero cost portfolio regression is negative, not positive as
expected under the null, at −0.21 with a t-statistic of −0.95. The 95% confidence
interval is from −0.63 to 0.22. While the 95% confidence interval contains positive
numbers up to 0.22, so that it is possible at that level of significance to fail to
reject the limits of arbitrage argument for differences up to that level, the Bayesian
posterior odds put the odds on the limits of arbitrage argument at only 0.20 for
small value stocks. That is, with even prior odds before the regression, the Bayesian
posterior odds inference is 5:1 ([1/0.20]:1) against the hypothesis that the high
residual variability small value stocks earn higher returns (as expected under the
null of the limits of arbitrage argument) versus the opposite. Put another way, even
someone who came to the data with a strong prior belief 5:1 in favor of the limits
of arbitrage argument for small value stocks (i.e., 83.3% probability that the limits
of arbitrage argument for small value stocks is true, 16.7% probability that it is
not) would be left indifferent after seeing the data, with equal parts belief for and
against the limits of arbitrage argument for small value stocks.7
Panel A of Table III presents results for growth stocks. We define “growth” as firms
in the bottom book to market quintile, using breakpoints for the full universe of
firms on the NYSE, AMEX, and NASDAQ. Firms are further allocated into quartile
sorts based on the magnitude of their estimated residual variability. Q1-Low (Q4-
High) is a portfolio holding firms with lowest (highest) residual variability. As with
small growth stocks, Panel A demonstrates that the poor performance in growth
stocks occurs in the highest residual variability portfolios.
7
It is possible that the correlation that we document between portfolios’ average return and standard
deviation is due to skewness in portfolio returns, where positive (negative) skewness leads to estimates
of mean and variance that are positively (negatively) correlated; We have tested for this possibility by
estimating the average portfolio return on the odd-numbered time-series observations and the standard
deviation on the even-numbered observations. This effectively generates estimates from independent
samples. We find that the results in this section as well as those below remain qualitatively unchanged.
THE LIMITS OF THE LIMITS OF ARBITRAGE 171
Table III. Post-formation results for the “growth” and “value” stock subsample: 1963–2007
We begin with the universe of firms traded on NYSE, AMEX, and NASDAQ with CRSP common
share codes 10 or 11. In Panel A we focus on firms in the lowest book-to-market quintile whereas
in Panel B we narrow the universe to firms in the highest book to market quintile. The portfolio
formation and resulting characteristics are defined in Table I.
The low returns to growth stocks are a phenomenon of the high residual vari-
ability portfolio. The high residual variability CAPM alpha is −1.05 percent per
month with a t-statistic of −4.01. The four-factor model cannot price high resid-
ual variability small growth stocks. The four-factor alpha of the highest residual
variability portfolio is −0.82 with a t-statistic of −3.83. The CAPM alpha of the
low residual variability portfolio is insignificantly different from zero, while the
four-factor alpha is positive and statistically significant, partly, it appears, because
low residual variability growth stocks earn good returns despite loading negatively
on the HML factor. That is, low residual variability growth firms covary strongly
with other growth stocks but do not suffer their same poor returns. High residual
variability growth stocks covary even more strongly (negatively) with HML but
perform much worse than expected under the four factor model given their large
market betas and SMB loadings.
The Sharpe Ratio of the highest residual variability subset of growth stocks is
−0.10 while the lowest residual variability growth stocks have a Sharpe Ratio of
0.30. The low Sharpe Ratio found in high residual variability growth stocks is both
a return and standard deviation effect. High residual variability growth stocks return
minus 28 basis points per month, on average, with a standard deviation of about
10% per month. The anomalously poor performing growth stocks comprise less
than 3% of the CRSP universe market capitalization.
Table II presents the results of zero cost portfolio CAPM regressions that are long
the high residual variability growth portfolio and short the low residual variability
growth portfolio. Under the null of the limits of arbitrage argument, the sign on
the growth regression alpha should be negative, since the negative alpha for growth
stocks should be strongest in the highest residual variability portfolio if the anomaly
is strongest where limits to arbitrage are highest. This is what we find. The CAPM
alpha of the zero cost portfolio regression is −1.01 with a t-statistic of −3.43. The
95% confidence interval is from −1.59 to −0.43. The posterior odds for the limits
of arbitrage argument are about 3,000 to 1 ([1/3.02 E-04]:1).
Panel B of Table III presents results for value stocks. These are firms which are
classified in the sorting month as falling in the top book to market quintile. The high
residual variability CAPM alpha is insignificant. There is no value stock anomaly
in high residual variability value stocks. The low residual variability CAPM alpha
is 40 basis points per month with a t-statistic of 3.24. The Sharpe ratio of the lowest
residual variability portfolio, 0.60, is nearly 40% higher than the Sharpe ratio of
the highest residual variability portfolio, 0.44. This is a standard deviation effect.
The high residual variability portfolio earns 24 basis points a month more than the
lowest residual variability portfolio but its standard deviation is almost twice as
high.
Table II presents the results of zero cost portfolio CAPM regressions that are long
the high residual variability value portfolio and short the low residual variability
THE LIMITS OF THE LIMITS OF ARBITRAGE 173
value portfolio. Under the null of the limits of arbitrage argument, the sign on the
value regression alpha should be positive, since the positive alpha generated by value
stocks should be strongest in the highest residual variability portfolio if the anomaly
is strongest where limits to arbitrage are highest. As for small value stocks, we do
not find support for the limits of arbitrage argument for value stocks. The CAPM
alpha of the zero cost portfolio regression is negative, not positive as expected under
the null, at −0.05 with a t-statistic of −0.22. The 95% confidence interval is from
−0.48 to 0.38. While the 95% confidence interval contains positive numbers up
to 0.38, so that it is possible at that level of significance to fail to reject the limits
of arbitrage argument for differences up to that level, the Bayesian posterior odds
put the odds on the limits of arbitrage argument at 0.70 for value stocks. That is,
with even prior odds before the regression, the Bayesian posterior odds inference
is 1.41:1 against the hypothesis that the high residual variability value stocks earn
higher returns (as expected under the null of the limits of arbitrage argument)
versus the opposite. Put another way, even someone who came to the data with
a prior belief 1.41:1 in favor of the limits of arbitrage argument for value stocks
(i.e., 70.5% probability that the limits of arbitrage argument for value stocks is true,
29.5% probability that it is not) would be left indifferent after seeing the data, with
equal parts belief for and against the limits of arbitrage argument for value stocks.
Panel A of Table IV presents results for recent winners. We define recent win-
ners as firms whose 11-month buy and hold return leading up to the month
prior to the formation period places them in the top quintile of price momentum.
Panel A of Table IV demonstrates that the abnormal return in recent winners is weak-
est in the highest residual variability environment. There is no CAPM anomaly for
recent winners in high residual variability portfolios: the high residual variability
alpha is an insignificant 0.14 basis points per month. The low residual variability
CAPM alpha, however, is 36 basis points per month with a t-statistic of 3.44. The
Sharpe Ratio of the highest residual variability subset of recent winners is 0.34
while the lowest residual variability recent winners have a Sharpe Ratio of 0.59.
This is a standard deviation effect. High residual variability recent winners return
6 basis points more than the lowest residual variability portfolio while the standard
deviation of the highest residual variability portfolio is 8.94 compared with 4.80
for the low residual variability portfolio. In Panel B, we restrict the sample to small
winners, firms that belong to the bottom size quintile on CRSP. None of conclusions
drawn are sensitive to conditioning on firm size.
Table II presents the results of zero cost portfolio CAPM regressions that are
long the high residual variability winner (and small winner) portfolio and short the
low residual variability winner (and small winner) portfolio. Under the null of the
174 A. BRAV ET AL.
Table IV. Post-formation results for the “winner” and “loser” stock subsample: 1963–2007
We begin with the universe of all firms traded on NYSE, AMEX, and NASDAQ with CRSP common
share codes 10 or 11. “Winners” (“Losers”) are determined by assigning firms to the highest (lowest)
price momentum quintile. Price momentum is measured using past one-year return skipping the
pre-formation month. In Panels B and D we further restrict the momentum samples to “Small” firms,
those whose market capitalization is in the bottom size quintile. The quartile portfolio formation and
resulting characteristics are defined in Table I.
limits of arbitrage argument, the sign on the winner and small winner regression
alphas should be positive, since the positive alpha generated by winner and small
winner stocks should be strongest in the highest residual variability portfolio if the
anomaly is strongest where limits to arbitrage are highest. As with small value and
value stocks, we do not find support for the limits of arbitrage argument for winner
or small winner stocks. The CAPM alphas of both zero cost portfolio regression
are negative, not positive as expected under the null, at −0.23 with a t-statistic of
−0.86 for winners, and −0.96 with a t-statistic of −3.59 for small winners. The
95% confidence interval for winners is from −0.74 to 0.29 and for small winners
is −1.48 to −0.44. While the 95% confidence interval for winners (but not small
winners) contains positive numbers up to 0.29, so that it is possible at that level
of significance to fail to reject the limits of arbitrage argument for differences up
to that level, the Bayesian posterior odds put the odds on the limits of arbitrage
argument for winners at 0.24. That is, with even prior odds before the regression,
the Bayesian posterior odds inference is 4.17:1 against the hypothesis that the high
residual variability value stocks earn higher returns (as expected under the null
of the limits of arbitrage argument) versus the opposite. The posterior odds for
small winners are 1.68 E-04, or more than 6,000 to 1 against the limits of arbitrage
argument.
Panel C of Table IV presents results for recent losers. Consistent with our results
on other negative anomalies, the worst alphas and Sharpe Ratio are in the highest
residual variability portfolio. When we further restrict the sample to small loser
firms in Panel D we find again that the conclusions are not sensitive to conditioning
on firm size. Table II reports that the posterior odds are overwhelmingly in favor
of the limits of arbitrage hypothesis. In unreported results we find that high limits
losers are disproportionately small growth firms, consistent with the evidence in
Table I, Panel B.8
In this section we extend our analysis and ask whether similar evidence exists
in event-study-based anomalous price reactions. In particular, we focus on the
well-known post-earnings announcement drift (Ball and Brown, 1968; Bernard and
Thomas, 1990), the evidence that stock prices tend to drift in the same direction as
the earnings news for three quarters post-event. We begin with the universe of all
firms traded on NYSE, AMEX, and NASDAQ with CRSP common share codes 10
8
In unreported analysis, we check whether results in this section are sensitive to industry concentra-
tion, namely, whether low or high idiosyncratic risk stocks tend to concentrate in specific industries.
We find that low limits to arbitrage stocks tend to have higher representation from industries such as
banking, utilities, and financials. However, when we remove these industries from the analysis in this
section and repeat our tests we find that our conclusions remain unchanged.
THE LIMITS OF THE LIMITS OF ARBITRAGE 177
D1 D2 D3 D4 D5 D6 D7 D8 D9 D10
Alpha 0.08 −0.04 −0.16 −0.14 −0.09 0.00 0.11 0.09 0.08 0.19
T-statistic 0.90 −0.46 −1.90 −1.83 −1.15 −0.02 1.37 1.16 1.05 2.24
Q1 Q2 Q3 Q4 Q5
Q1 Q2 Q3 Q4 Q5
portfolios and calculate value weight results for the ensuing three months. We
rebalance these portfolios quarterly. The regression alphas are presented in Panel C
of Table V. The first two rows (“Low IV Alpha”) provide quintile SUE portfolio
alphas and t-statistics for low residual variability stocks. The next two rows (“High
IV Alpha”) provide alphas and t-statistics for quintile SUE portfolios formed using
high residual variability stocks.
Negative drift for extreme negative SUE portfolio is significant in both the high
volatility portfolio with a negative 87 basis point alpha (t-statistic = −4.72) and
the low volatility portfolio yields a marginally significant alpha of negative 35
basis points (t-statistic = −3.25). This evidence is consistent with the predictions
regarding overvaluation that we have discussed earlier. However, we can also see
that the positive drift subsequent to positive earnings news is confined entirely to
low residual variability portfolios, whereas the high residual variability portfolio
alphas are all insignificantly different from zero. The alpha in the extreme positive
SUE portfolio is large and significant only in the low volatility portfolio with a 62
basis point alpha (t-statistic = 5.42). The corresponding high volatility portfolio
yields an insignificant alpha of negative 2 basis points (t-statistic = −0.12).
Table II presents the results of zero cost portfolio CAPM regressions that are
long the high residual variability portfolio of positive earnings surprises for small
firms and short the low residual portfolio of positive earnings surprises for small
firms. Under the null of the limits of arbitrage argument, the sign on the regression
alpha should be positive, since the positive alpha generated by the portfolio of
positive earnings surprises for small firms should be strongest in the highest residual
variability portfolio if the anomaly is strongest where limits to arbitrage are highest.
We do not find support for the limits of arbitrage argument for the portfolios of
positive earnings surprises for small firms. The CAPM alpha of the zero cost
portfolio regression is negative, not positive as expected under the null, at −0.64
with a t-statistic of −3.20. The 95% confidence interval is from −1.02 to −0.25.
The Bayesian posterior odds put the odds on the limits of arbitrage argument at only
0.20 for small value stocks. That is, with even prior odds before the regression, the
Bayesian posterior odds inference is 355:1 ([1/2.98 E-03]:1) against the hypothesis
that the high residual variability portfolio earns higher returns (as expected under
the null of the limits of arbitrage argument) versus the opposite. As with our earlier
results, the evidence is consistent with the limits of arbitrage for the negative
earnings surprises for small firms, with a CAPM alpha of −0.58, a t-statistic of
−2.75, a 95% confidence interval of −1.00 to −0.17, and Bayesian posterior odds
of about 1453:1 in favor of the limits of arbitrage argument.10
10
The evidence from recent work on the “accrual anomaly” (see Sloan, 1996) by Lev and Nissim
(2006) indicates that our conclusions extend beyond the “anomalies” studied in this paper. Lev and
Nissim (2006) show that the inverse relation between accounting accruals and subsequent returns is
180 A. BRAV ET AL.
The tests presented in Section 4 were based on predictions from the behavioral
literature linking residual variability (and other correlated firm characteristics) to
the magnitude of observed anomalies. Consider again the description from Barberis
and Thaler (2003):
Noise trader risk . . . is the risk that the mispricing being exploited by the arbitrageur worsens
in the short run. . . . [T]he arbitrageur [] faces the risk that the pessimistic investors causing
Ford to be undervalued in the first place become even more pessimistic, lower its price even
further. Once one has granted the possibility that a security’s price can be different from its
fundamental value, then one must also grant the possibility that future price movements will
increase the divergence.
Under this hypothesis, the superiority of value stocks should be highest for
stocks exhibiting “loser” momentum trends since arbitrageurs, who are aware of
these momentum characteristics, are presumably reluctant to purchase value stocks
that are more likely to continue to underperform in the short run. Similarly, the
underperformance of growth stocks should be stronger for firms exhibiting recent
high price appreciation, namely, momentum “winners,” because arbitrageurs will
not want to bet against growth stocks that are more likely to continue to appreciate
in price.
Table VI presents results for these tests. We form four portfolios by sorting
firms based on their pre-formation book-to-market and momentum characteristics.
The sample is the universe of all firms traded on NYSE, AMEX, and NASDAQ
with CRSP common share codes 10 or 11, further constrained to firms in the
bottom quintile by firm size.11 We form four portfolios requiring that each be al-
located a minimum of 50 firms in the formation period. The portfolios are value
weighted and rebalanced quarterly. The first portfolio, ‘Loser-Value,’ holds firms
in the bottom momentum quintile and highest book-to-market quintile. The second
portfolio, ‘Loser-Growth,’ holds stocks in the bottom quintiles by both momen-
tum and book-to-market. The third portfolio, ‘Winner-Value,’ holds stocks in the
highest momentum and book-to-market quintiles. The fourth portfolio, ‘Winner-
Growth,’ holds stocks in the highest momentum quintile and bottom book-to-market
confined to firms with high accruals that, ex-post, tend to underperform standard size and book to
market benchmarks (see their Table I and the discussion in Section 6). There is no reliable evidence
of undervaluation. Lev and Nissim (2006) also show that high accrual firms are characterized by
high residual variability, low market capitalizations, low book to market ratios, low prices, and
reduced likelihood of institutional ownership. These firms have precisely the same characteristics of
overvalued firms that we find in high limits to arbitrage environments.
11
We focus on the bottom size quintile since both value and momentum premia are larger for this
subset of stocks. We have conducted the same analysis for the full universe and the results remain
qualitatively unchanged.
THE LIMITS OF THE LIMITS OF ARBITRAGE 181
while the ‘Winner-Growth’ portfolio has a CAPM alpha of 0.07 (t-statistic 0.24),
the opposite of the outcome expected under the noise trader momentum risk null
hypothesis.
In our final set of tests, we ask whether there is a factor premium for size, book-to-
market, and momentum in portfolios comprised of the lowest residual variability
stocks. The lowest residual variability stocks not only have low idiosyncratic risk
but also have high median prices, high institutional holdings, a larger number of
analysts, considerable liquidity, high dividend yields, and comprise a large part of
the market capitalization of the entire market. It is difficult to argue that limits of
arbitrage are meaningful in these stocks. If the factor premiums for size, book-to-
market, and momentum were driven entirely by irrationality then we would expect
very little covariation with these factors in the lowest residual variability portfolios.
Our previous results already shed some light on this hypothesis: even our lowest
limits to arbitrage portfolios load strongly on the SMB and HML factors. To test
this more directly, Table VII reports the results of zero cost portfolio regressions
where the relevant zero cost portfolio is long one low limits portfolio and short
another low limits portfolio. Panel A reports the results of regressions of the returns
to the zero cost portfolio that is long low limits small stocks and short low limits
large stocks. Each low limit portfolio consists of firms whose residual standard
deviation is in the bottom quartile of the respective distribution of pre formation
residual standard deviations. The first regression in Panel A is of the returns to that
zero cost portfolio on the other known sources of covariation in stocks returns, i.e.,
the three other factors, excluding SMB. The low limits small-minus-big portfolio
loads 0.28 on HML with a t-statistic of 3.36 but does not load significantly on
RMRF or MOM. There remains an unexplained alpha of 34 basis points per month
with a t-statistic of 1.96. This suggests that there is a size premium even in the
lowest limits to arbitrage portfolios. The next regression of Panel A adds SMB to
the zero cost regression. The loading on SMB is 0.94 with a t-statistic of 9.75.
The alpha then declines to an insignificant 10 basis points. Table VIII, panel A,
presents the same results for the high limits portfolios. While high limits firms do
load more strongly on SMB, there remains a large unexplained negative alpha of
−0.71 (t-statistic −3.71), apparently because the high limits small perform much
worse than would be predicted by their large loading on SMB.
Panel B of Table VII reports the results of regressions on the zero cost portfolio
that is long low limits value stocks and short low limits growth stocks. The first
regression in Panel B is on the three other factors excluding HML. The low lim-
its high-minus-low portfolio loads negatively on RMRF, positively on SMB, and
THE LIMITS OF THE LIMITS OF ARBITRAGE 183
panel C, presents the same results for the high limits portfolios. While high limits
small firms do load more strongly on HML, there remains a large unexplained
alpha of 1.49 (t-statistic 4.70), again, apparently because the high limits growth
firms perform much worse than would be predicted by the zero cost portfolio’s
loading on HML.
Panel D of Table VII reports the results of regressions on the zero cost portfolio
that is long low limits recent winners and short low limits recent losers. The first
regression in Panel D is on the three other factors excluding MOM. The low limits
winners-minus-losers portfolio loads slightly negatively on RMRF, SMB, and HML.
There remains an unexplained alpha of 66 basis points a month with a t-statistic
of 2.55. This suggests that there is a momentum premium even in the lowest limits
to arbitrage portfolios. The next regression of Panel D adds MOM to the zero cost
regression. The loading on MOM is 1.24 with a t-statistic of 30.22. The alpha
is now negative at 0.52 with a t-statistic of −4.71, apparently because the zero
cost portfolio, while moving strongly with MOM, does not have a sufficiently low
return since low limits to arbitrage losers do not underperform. Table VIII, panel
D, presents the same results for the high limits portfolios. High limits firms do not
load more strongly on MOM, and there remains a large unexplained positive alpha
of 0.84 (t-statistic 3.76), apparently because the high limits losers firms perform
much worse than would be predicted by the zero cost portfolio’s loading on MOM.
Table VII, panel E, presents the results for only low limits small winner minus
small losers. The alpha now becomes positive, apparently because low limits small
winners perform much better than would be predicted by the zero cost portfolio’s
loading on MOM. Table VIII, panel E, presents the same results for the high limits
portfolios. The results are qualitatively the same, though the alpha is even larger,
apparently because high limits losers perform much worse than would be predicted
by the zero cost portfolio’s loading on MOM.
7. Conclusions
(low returns to growth stocks, recent losers, and negative earnings surprises). We
find no support, however, for the noise trader momentum risk version of the limits
of arbitrage argument. Taken together, these results are a success for behavioral
explanations of overvaluation anomalies but present a serious challenge to develop
new behavioral explanations for the existence of undervaluation anomalies.
References
Amihud, Y. (2002) Illiquidity and stock returns: cross-section and time-series effects, Journal of
Financial Markets 5, 31–56.
Asness, S. C. (1997) The interaction of value and momentum strategies, Financial Analysts Journal,
March/April, 29–36.
Ali, A., Hwang, L-S., and Trombley, M. A. (2003) Arbitrage risk and the book-to-market anomaly,
Journal of Financial Economics 69, 355–373.
Ang, A., Hodrick R., Xing Y., and Zhang X. (2006) The cross-section of volatility and expected
returns, Journal of Finance 61, 259–299.
Ball, R. and Brown, P. (1968) An empirical evaluation of accounting numbers, Journal of Accounting
Research 6, 159–178.
Banz, R. (1981) The relationship between return and market value of common stocks, Journal of
Financial Economics 9, 3–18.
Barberis, N. and Huang M. (2008) Stocks as lotteries: the implications of probability weighting for
security prices, American Economic Review 98, 2066–2100.
Barberis, N., Shleifer, A., and Vishny, R. W. (1998) A model of investor sentiment, Journal of
Financial Economics 49, 307–343.
Barberis, N. and Thaler, R. (2003). A survey of behavioral finance, in: G. M. Constantinides,
M. Harris, and R. M. Stulz (eds.), Handbook of the Economics of Finance, Elsevier, First edition,
Volume 1, Chapter 18, pp. 1053–1128.
Bernard, V. L. and Thomas, J. K. (1990) Evidence that stock prices do not fully reflect the im-
plications of current earnings for future earnings, Journal of Accounting and Economics 13, 305–
340.
Brav, A. and Heaton, J. B. (2002) Competing theories of financial anomalies, Review of Financial
Studies 15, 575–606.
Campbell, J. Y., Lettau, M., Malkiel, B. G., and Xu, Y. (2001) Have individual stocks become more
volatile? An empirical exploration of idiosyncratic risk, Journal of Finance 56, 1–43.
Chan, L. K. C., Jegadeesh, N., and Lakonishok, J. (1996) Momentum strategies, Journal of Finance
51, 1681–1713.
Chordia, T., Goyal, A., Sadka, G., Sadka, R., and Shivakumar, L. (2006) Liquidity and the post-
earnings-announcement-drift, unpublished working paper, University of Washington.
Cross, R. (1982) The Duhem-Quine thesis, Lakatos and the appraisal of theories in macroeconomics,
The Economic Journal 92, 320–340.
Daniel, K., Hirshleifer, D., and Subrahmanyam, A. (1998) Investor psychology and security market
under- and overreactions, Journal of Finance 53, 1839–1886.
DeLong, J. B., Shleifer, A., Summers, L. H., and Waldmann, R. J. (1990) Noise trader risk in financial
markets, Journal of Political Economy 98, 703–738.
Fama, E. F. (1965) The behavior of stock market prices, Journal of Business 38, 34–105.
Fama, E. F. (1991) Efficient capital markets: II, Journal of Finance 46, 1575–1617.
THE LIMITS OF THE LIMITS OF ARBITRAGE 187
Fama, E. F. and French, K. (1993) Common risk factors in the returns on stocks and bonds, Journal
of Financial Economics 33, 3–56.
Fama, E. F. and French, K. (1995) Size and book-to-market factors in earnings and returns, Journal
of Finance 50, 131–155.
Friedman, M. (1953) The case for flexible exchange rates, in: Essays in Positive Economics, University
of Chicago Press, Chicago.
Geczy, C., Musto D. K., and Reed, A. V. (2002) Stocks are special too: an analysis of the equity
lending market, Journal of Financial Economics 66, 241–269.
Grossman, S. J. and Stiglitz, J. E. (1980) On the impossibility of informationally efficient markets,
American Economic Review 70, 393–408.
Hoberg, G. and Prabhala, N. (2009) Disappearing dividends, catering, and risk, Review of Financial
Studies 22, 79–116.
Hong, H. and Stein, J. (1999) A unified theory of underreaction, momentum trading and overreaction
in asset markets, Journal of Finance 54, 2143–2184.
Kogan, L., Ross, S. A., Wang, J., and Westerfield, M. M. (2006) The price impact and survival of
irrational traders, Journal of Finance 61, 195–229.
Lakonishok, J., Shleifer, A., and Vishny, R. (1994) Contrarian investment, extrapolation, and risk,
Journal of Finance 49, 1541–1578.
Loughran, T. and Ritter, J. R. (2000) Uniformly least powerful tests of market efficiency, Journal of
Financial Economics 55, 361–389.
Lev, B. and Nissim, D. (2006) The persistence of the accruals anomaly, Contemporary Accounting
Research 23, 1–34.
Loewenstein, M. and Willard, G. A. (2006) The limits of investor behavior, Journal of Finance 61,
231–258.
Mashruwala, C., Rajgopal, S., and Shevlin, T. (2006) Why is the accrual anomaly not arbitraged
away? The role of idiosyncratic risk and transaction costs, Journal of Accounting and Economics 42,
3–33.
Mendenhall, R. (2004) Arbitrage risk and post-earnings-announcement drift, Journal of Business 77,
875–894.
Pontiff, J. (1996) Costly arbitrage: evidence from closed-end funds, The Quarterly Journal of Eco-
nomics 111, 1135–1151.
Pontiff, J. (2006) Costly arbitrage and the myth of idiosyncratic risk, Journal of Accounting and
Economics 42, 35–52.
Ritter, J. R. (1991) The long-run performance of initial public offerings, Journal of Finance 46, 3–27.
Shleifer, A. and Vishny, R. (1997) The limits of arbitrage, Journal of Finance 52, 35–55.
Shleifer, A. (2000) Inefficient Markets, Oxford University Press.
Sloan, R. (1996) Do stock prices fully reflect information in accruals and cash flows about future
earnings?, The Accounting Review 71, 289–315.
Wurgler, J., and Zhuravskaya, E. V. (2002) Does arbitrage flatten demand curves for stocks?, Journal
of Business 75, 583–608.
Zellner, A. and Siow, A. (1979) On posterior odds ratios for sharp null hypotheses and one-sided
alternatives, H.G.B. Alexander Research Foundation, Graduate School of Business, University of
Chicago.
Zhang, X. F. (2006) Information uncertainty and stock returns, Journal of Finance 61, 105–136.